Radiation damage has always limited resolution in biological imaging using electrons or X-rays2
. With the recent invention of the femtosecond X-ray laser, an opportunity has arisen to break the nexus between radiation dose and spatial resolution. It has been proposed that femtosecond X-ray pulses can be used to outrun even the fastest damage processes by using single pulses so brief that they terminate before the manifestation of damage to the sample6
. Experiments at the FLASH free-electron laser (FEL), Germany, confirmed the feasibility of ‘diffraction before destruction’ at resolution lengths down to 60 Å on test samples fixed on silicon nitride membranes7
. It was predicted that the irradiance (or power density) of focused pulses from a hard-X-ray FEL such as the Linac Coherent Light Source (LCLS), USA, would be sufficient to produce diffraction patterns at near-atomic resolution6
We demonstrate here that this notion of diffraction before destruction operates at subnanometre resolution, using the membrane protein photosystem I as a model system, and establish an approach to structure determination based on X-ray diffraction data from a stream of nanocrystals6,8
. Membrane proteins have a central role in the functioning of cells and viruses, yet our knowledge of the structure and dynamics responsible for their functioning remains limited. Photosystem I is a large membrane protein complex (1-MD a molecular mass, 36 proteins, 381 cofactors) that acts as a biosolar energy converter in the process of oxygenic photosynthesis. Its crystals display the symmetry of space group P
, with unit-cell parameters a
= 281 Å and c
= 165 Å, and consist of 78% solvent by volume. We show that diffraction data can be recorded from these fragile protein nanocrystals before destruction occurs. Furthermore, we demonstrate that structure factors can be extracted from the ‘partial’ reflections of tens of thousands of single-crystal diffraction snapshots, showing that interpretable high-quality, three-dimensional (3D) structure factor data can be obtained from a suspension of submicrometre crystals.
Our experimental set-up ( and Methods) records single-crystal diffraction data from a stream of crystals carried in a 4-μm-diameter, continuous liquid water jet9
that flows across the focused LCLS X-ray beam in vacuum at 10 μl min−1
. In contrast to cryo-electron microscopy10,11
or standard crystallography on microcrystals3
, which require cryogenic cooling, these data were collected on fully hydrated, 3D nanocrystals. The crystal located in the interaction region when an X-ray pulse arrives gives rise to a diffraction pattern that is detected on a set of two low-noise, X-ray p–n junction charge-coupled device (pnCCD) modules12
and read out before the arrival of the next pulse at the FEL repetition rate of 30 Hz, or 1,800 patterns per minute. The photon energy of the X-ray pulses was 1.8 keV (6.9-Å wavelength), with more than 1012
photons per pulse at the sample and pulse durations of 10, 70, and 200 fs (ref. 13
). An X-ray fluence of 900 J cm−2
was achieved by focusing the FEL beam to a full-width at half-maximum of 7 μm, corresponding to a sample dose of up to 700 MGy per pulse (calculated using the program RADDOSE14
) and a peak power density in excess of 1016
at 70-fs duration. In contrast, the typical tolerable dose in conventional X-ray experiments is only about 30 MGy (ref. 1
). A single LCLS X-ray pulse destroys any solid material placed in this focus, but the stream replenishes the vaporized sample before the next pulse.
The front detector module, located close to the interaction region, recorded high-angle diffraction to a resolution of 8.5 Å, whereas the rear module intersected diffraction at resolutions in the range of 4,000 to 100 Å. We observed diffraction from crystals smaller than ten unit cells on a side, as determined by examining the data recorded on the rear pnCCDs (). A crystal with a side length of N
unit cells gives rise to diffraction features that are finer by a factor of 1/N
than the Bragg spacing (that is, with N
− 2 fringes between neighbouring Bragg peaks), providing a simple way to determine the projected size of the nanocrystal. Images of crystal shapes obtained using an iterative phase retrieval method15,16
are shown in . The 3D Fourier transform of the crystal shape is repeated on every reciprocal lattice point. However, the diffraction condition for lattice points is usually not exactly satisfied, so each recorded Bragg spot represents a particular ‘slice’ of the Ewald sphere through the shape transform, giving a variety of Bragg spot profiles in a pattern; these are apparent in . The sum of counts in each Bragg spot underestimates the underlying structure factor square modulus, representing a partial reflection.
shows strong single-crystal diffraction to the highest angles of the front detector. The nanocrystal shape transform is also apparent in many patterns at the high angles detected by the front detector, giving significant measured intensities between Bragg peaks as is noticeable in Supplementary Fig. 3a
. These mid-Bragg intensities oversample the molecular transform, providing a potential route to phasing of the pattern17,18
Diffraction intensities and electron density of photosystem I
In conventional crystallography, the ‘full’ Bragg reflection is determined to high precision, for example by integrating counts as the crystal is rotated such that these reflections pass through the diffraction condition. By indexing individual patterns and then summing counts in all partial reflections for each index, we performed a Monte Carlo integration over the reciprocal-space volume of the Bragg reflection and the distribution of crystal shapes and orientations and variations in the X-ray pulse fluence. The result of this procedure converges to the square of the structure factor moduli18
. We found that over 13% of diffraction patterns with ten or more spots could be consistently indexed using the programs MOSFLM19
(Methods). Merged intensities at 70-fs pulse duration are presented as a precession-style image of the -zone axis in (see also Supplementary Figs 3 and 4
). We tested the reliability of this approach by comparing the LCLS merged data with data collected at 100 K with 12.4-keV synchrotron radiation from a single crystal of photosystem I cryopreserved in 2 M sucrose. These data sets show good agreement, with a difference metric, Riso
, of 22.1% computed over the entire resolution range and of less than 13% in the middle resolution shells; see Supplementary Table 1
for detailed statistics.
To complete our proof of principle, we conducted a rigid-body refinement of the published photosystem I structure (Protein Data Bank ID, 1JB0) against the nanocrystal structure factors, yielding R/Rfree = 0.25/0.23. A representative region of the 2mFo − DFc electron density map at 8.5 Å (Methods) from the LCLS data set is shown in . This map shows the details expected at this resolution, including transmembrane helices, membrane extrinsic features and some loop structures. For comparison, the electron density refined from the 12.4-keV, single-crystal data set truncated to a resolution of 8.5 Å is given in .
The dose of 700 MGy corresponds to a K-shell photoabsorption of 3% of all carbon atoms in the protein. This energy is subsequently released by photoionization and Auger decay, followed by a cascade of lower-energy electrons caused by secondary ionizations, taking place on the 10–100-fs timescale21
. Using a model of the plasma dynamics22,23
, we calculated that by the end of a 100-fs pulse each atom of the crystal was ionized once, on average, and that motion of nuclei had begun. This is expected to give rise to a decrease in Bragg amplitudes, similar to an increase in a Debye–Waller temperature factor24
. We studied the effects of the initial ionization damage on the diffraction of photosystem I nanocrystals by collecting a series of data sets at pulse durations of 10, 70 and 200 fs. The 10-fs pulses were produced with lower pulse energy: ~10% of the total number of photons of the longer pulses13
, or a 70-MGy dose. Plots of the scattering strength of the crystals versus resolution, generated by selecting and summing Bragg spots from more than 66,000 patterns for each of the three pulse durations measured, are shown in . The 10- and 70-fs traces are very similar, indicating that these pulses are short enough to overcome radiation damage at the observed resolution, 8.5 Å. For 200-fs pulses, there is a decrease in scattering strength at resolutions beyond 25 Å, indicating disordering on this longer timescale. The highest-resolution Bragg peaks for the 200-fs pulses were not broadened or shifted relative to the short-duration data sets, which indicates there was no strain or expansion of the lattice, respectively.
Pulse-duration dependence of diffraction intensities
Our next step is to improve resolution by using shorter-wavelength X-rays. Resolution may ultimately be limited by X-ray pulse fluence, the ultrafast radiation damage and the intrinsic disorder within the nanocrystals themselves. Recent experiments21
at LCLS indicate a brief saturation of the X-ray photoabsorption of atoms in a tightly focused pulse, resulting in a decrease in photoionization damage on a 20-fs timescale without a reduction in the scattering cross-sections that give rise to the diffraction pattern22
. Planned beamlines at LCLS aim to achieve up to a 105
-fold increase in pulse irradiance by tighter focusing, allowing data collection with low-fluence, 10-fs pulses or pulses of even shorter duration25
. This provides a route to further reducing radiation damage and may allow measurements on even smaller nanocrystals, down to a single unit cell6
(that is, a single molecule). As this limit is approached, the ordering of the nanocrystals will become increasingly irrelevant, as each crystal may be treated as a single object and the ‘disorder’ that conventionally leads to reduced resolution will simply manifest itself as shot-to-shot variability, providing information about not just the average structure but also the range of dynamically accessible conformations.
Data are collected on fully hydrated nanocrystals without cryogenic cooling. We expect that the results presented here will open new avenues for crystallography using X-ray laser pulses that are so short that only negligible X-ray-induced radiation damage occurs during data collection. Significant improvements in sample utilization are expected by exploiting higher X-ray repetition rates or by slowing the liquid flow. For example, the generation, using inkjet technologies, of liquid droplets at a rate that matches the LCLS X-ray pulses would dramatically decrease the total required sample volume by a factor of 25,000, meaning that less than 0.4 μl of nanocrystal suspension would be needed in our particular case, of photosystem I. Further efficiency gains would result from indexing and merging a greater proportion of patterns into the 3D data set, which may be achieved by applying methods for merging continuous diffraction patterns of single molecules26,27
or by using ‘post-refinement’28
to obtain accurate structure factor estimates from fewer diffraction patterns. These methods will also remove the twinning ambiguity that exists in our current indexing scheme. Our method also has potential application to the study of chemical reactions, such as the processes in photosynthesis or enzymatic reactions.